Estimating the "Wrong" Graphical Model: Benefits in the Computation-Limited Setting